基于L1范数的二维鉴别局部保留投影的最大间距准则方法  

L1-Norm Based Maximum Margin Criterion of Two-dimensional Discriminant Locality Preserving Projection

作　　者：谢玉凯 卢桂馥[1] 宣东东 

机构地区：[1]安徽工程大学计算机与信息学院,安徽芜湖241000

出　　处：《长江大学学报（自然科学版）》2017年第9期38-42,78,共6页Journal of Yangtze University(Natural Science Edition)

基　　金：国家自然科学基金项目(61572033;71371012);安徽高校自然科学研究项目(重大项目)(KJ2015ZD08);教育部人文社会科学规划项目(13YJA630098)

摘　　要：为避免基于L1范数的鉴别局部保留投影(DLPP-L1)方法中图像数据向量化引起的矩阵计算的不稳定问题,提出了一种基于L1范数的二维鉴别局部保留投影的最大间距准则方法 (2DDLPP-L1/MMC),目的是寻找一组最优投影向量使得投影变换后特征空间的类间离散度最大而类内离散度最小。2DDLPP-L1/MMC方法直接基于图像矩阵并且利用了L1范数对野值处理的稳健性以及最大间距准则所具有的优势,降低了计算量。人脸数据库上的试验结果表明,该方法具有较强的鲁棒性和较高的识别率。When L1-norm based Discriminant Locality Preserving Projections for dimensionality reduction is applied,the image matrices needs to be concatenated into high-deminsion vectors which cause instability.To solve this problem,L1-norm based Maximum Margin Criterion of Two-dimensional Discriminant Locality Preserving Projection is presented.The proposed approach,which makes full use of maximum margin criterion and strong robustness of L1-norm to outliers,intends to maximize the inter-class scatter while simultaneously minimizing the intra-class scatter after projection.Experiments on face image databases demonstrate that the performance and robustness of 2DDLPP-L1/MMC are better than the related methods.

关 键 词：鉴别局部保留投影 L1范数 最大间距准则 鲁棒性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]